What happened
On 29 October Mistral AI, the French artificial intelligence company, released Mistral Large 4. Internally they call it Le Chonk, and the nickname fits: it has one trillion parameters, although only a part of them, about 52 billion, works at each step. This is known as mixture of experts, and it is why the model is faster and cheaper than its size suggests.
Two things make it different. The first is that it was trained from scratch on European infrastructure. The second is that the weights are open: anyone can download the model from Hugging Face and run it on their own servers.
If you do not want the hassle, you can use it like any other model, through an API, on Mistral's platform, on OpenRouter, on Microsoft Foundry inside Azure or on NVIDIA's services. It accepts text and images and Mistral says it holds its own in more than 160 languages.
The numbers worth knowing
We have checked these figures against Mistral's announcement and coverage from several specialist outlets:
- API price: 0.70 dollars per million input tokens, 0.07 if the input is cached and 2.10 per million output tokens.
- License: modified MIT. You can use it, adapt it and put it in your products. Only if you sell the model itself as a service to third parties do you need a separate commercial license.
- Performance: it reaches 48 % on Artificial Analysis's coding agent index. Claude Opus 5.5 sits at 62 %. It is strong, but it is not the best.
- Safety: it refuses 96.6 % of malicious requests in the tests Mistral publishes.
- Hosting it yourself: the standard version needs eight latest generation NVIDIA GPUs. Compressed, it drops to four. This is not a model for the office computer.
Why it matters
When we talk to companies about AI, the question that stalls projects most is not price. It is this one: where does my data go? Patient records, payroll, contracts, customer data. Many people would rather not upload that to a US service, and the AI Act and the GDPR give them reasons to think twice.
Until now the answer was awkward. The open models you could host yourself were clearly worse, or came from Chinese labs, which does not reassure some clients either. Large 4 offers a third way: a large model, from a European provider, that you can use through an API or run in your own cloud.
Price matters too. A few days ago we explained that GPT-6.1 Terra costs 1 dollar input and 5 dollars output per million tokens. Large 4 stays at 0.70 and 2.10. They are not in the same league for every task, but for many of the things a small business does the quality gap goes unnoticed and the cost gap does not.
What it means for a small business
Let's be realistic: almost no small business is going to set up a server with eight GPUs to host this model. That machine costs far more than you would save on the API, unless volume is huge and constant. The open weights matter mostly for what they allow your providers to do.
In practice we see three cases where Large 4 fits well:
Sectors with sensitive data. Clinics, law firms, accountants or HR companies that want to use AI on internal documents and need to be able to say, in writing, that a European provider processes that data.
Volume tasks. Summarising emails, classifying tickets, extracting data from documents or answering queries in several languages. There, a decent and cheap model beats a brilliant and expensive one.
Reducing dependence. If all your automation depends on a single provider, any price rise or change of terms hits you squarely. Having a second tested model is cheap insurance.
Where we would not put it yet is in complex agents that write code or make chained decisions. For that, the top models still perform clearly better, and the figures above confirm it.
What I would do this week
First, sort your tasks by sensitivity. List what AI already does in your company, or what you would like it to do, and mark which parts touch personal or confidential data.
Second, test Large 4 on one of those tasks. Take twenty or thirty real cases and compare it with what you use today: answer quality, time and cost. If you work with tools like n8n or Make, switching models usually takes minutes, and we have already explained how to choose between them.
Third, ask your providers. If your CRM, your chatbot or your management software uses AI, ask them which model sits behind it and where the data is processed. With Large 4 on the market, "there is no European alternative" is no longer a valid answer.
Fourth, do not migrate everything at once. The sensible approach is a mix: the European model for sensitive and repetitive work, and a top model only where quality justifies it.
Deciding which model goes where, and under what data terms, is exactly the work we do as an external CTO.
If you want us to look at it with you
We design AI agents and automations for small businesses, choosing the model that fits each task, not the one that is trending. We also build custom software when AI needs to be integrated with your systems.
Book a thirty-minute call. We will review with you what data your company handles, which tasks could move to a European model like Large 4 and how much you would save. No commitment.
Sources
This article draws on the following publications, consulted on 2 November 2026:
Frequently asked questions
It is an artificial intelligence model from French company Mistral AI, released on 29 October 2026. It has one trillion parameters, of which it uses about 52 billion at each step, accepts text and images and was trained from scratch on European infrastructure.
On Mistral's API it costs 0.70 dollars per million input tokens, 0.07 dollars if the input is cached and 2.10 dollars per million output tokens. The weights can be downloaded for free, but hosting it yourself requires servers with several high end GPUs.
Yes, the modified MIT license allows you to use it, adapt it and integrate it into your own products. You only need a separate commercial license if your business consists of selling the model itself as a service to third parties.
Not at everything. On Artificial Analysis's coding agent index it scores 48 %, against 62 % for Claude Opus 5.5. For volume tasks such as summarising, classifying or extracting data it is usually enough and cheaper; for complex agents the top models still perform better.
It can help, because it lets you work with a European provider or host the model in your own cloud. But the model alone does not guarantee compliance: you still need a data processing agreement, to know where the data is processed and to apply the measures the regulation requires.
In practice, no. The standard version needs eight latest generation NVIDIA GPUs and the compressed one, four. For a small business the sensible option is to use it through an API or a European cloud provider.